LEAKAGE-AWARE MULTI-SLICE MRI-BASED DEEP LEARNING FRAMEWORK FOR DEMENTIA STAGE CLASSIFICATION

Abstract

Dementia is a cognitive disorder characterized by a persistent reduction in memory, reasoning, and functional skills . Accurate and early diagnosis plays a crucial role in devising a successful treatment program . This report explores a deep learning approach for dementia classification based on multi-class MRI images utilizing the OASIS-2 dataset . In the initial stage, different CNN models, i.e., VGG16 , ResNet50 , DenseNet121 , and InceptionV3 , are trained on 2D MRI slices. While these models exhibit high accuracy, a careful examination suggests that leakage issues had caused in flated accuracy results due to erroneous slicing techniques that resulted in shared slices between training and validation sets from the same patients . The current study introduces a leakage-aware evaluation technique based on a data split according to patient level. An additional multi-slice data fusion process is also developed to analyze MRI slices during inference time and extract more anatomical details from brain regions. In the ultimate model, MRI data are fused with ten clinical variables (MMSE, nWBV, Age, CDR, eTIV, ASF, SES, EDUC, Sex, and MR Delay) using multi modal EfficientNetV2B0 architecture. Class imbalance issues are addressed by reducing dementia categories to three, namely Non-Demented, Mild-Demented, and Demented. The developed model attains a 70.48% accuracy rate, which is a fair indication of the model’s generalization capability, unlike the inflated baseline accuracies. The model performs satisfactorily in identifying Non-Demented and Demented individuals; however, it struggles with Mild-Demented diagnoses. In conclusion, this report illustrates the significance of applying a thorough evaluation technique for clinical deep learning models . Combining leakage-free training and multi-slice multimodal fusion improves dementia classification accu racy.